Upload
hoangxuyen
View
216
Download
2
Embed Size (px)
Citation preview
ORIGINAL ARTICLE
ARMin: a robot for patient-cooperative arm therapy
Tobias Nef Æ Matjaz Mihelj Æ Robert Riener
Received: 22 December 2006 / Accepted: 1 July 2007
� International Federation for Medical and Biological Engineering 2007
Abstract Task-oriented, repetitive and intensive arm
training can enhance arm rehabilitation in patients with
paralyzed upper extremities due to lesions of the central
nervous system. There is evidence that the training duration
is a key factor for the therapy progress. Robot-supported
therapy can improve the rehabilitation allowing more
intensive training. This paper presents the kinematics, the
control and the therapy modes of the arm therapy robot
ARMin. It is a haptic display with semi-exoskeleton
kinematics with four active and two passive degrees of
freedom. Equipped with position, force and torque sensors
the device can deliver patient-cooperative arm therapy
taking into account the activity of the patient and sup-
porting him/her only as much as needed. The haptic display
is combined with an audiovisual display that is used to
present the movement and the movement task to the
patient. It is assumed that the patient-cooperative therapy
approach combined with a multimodal display can increase
the patient’s motivation and activity and, therefore, the
therapeutic progress.
Keywords Robotics � Haptic device � Exoskeleton �Rehabilitation � Arm therapy
1 Introduction
1.1 Clinical background and rationale for arm therapy
Patients with paralysed upper extremities due to lesions of
the central nervous system, e.g. after stroke, traumatic
brain or spinal cord injury, often receive arm therapy. The
goal of this therapy is to recover motor function, improve
movement co-ordination, learn new motion strategies, so
called trick movements, and to prevent secondary com-
plications, such as muscle atrophy, osteoporosis, joint
degeneration and spasticity.
Several studies prove that arm therapy has positive
effects on the rehabilitation progress of stroke patients
(see [25] for review). Several groups observed that longer
daily training sessions and longer overall training periods
have a positive effect on the motor function of the arm
[18, 19, 33]. In a meta-analysis comprising nine controlled
studies with 1,051 stroke patients, Kwakkel et al. [17]
showed that increased training intensity yields moderate
positive effects on neuromuscular function and activities
of daily living (ADL). The study did not distinguish
between upper and lower extremities. The conclusion that
the rehabilitation progress depends on training intensity
and training duration supports the application of robot-
aided arm therapy.
There is further evidence that machine-delivered thera-
pies can enhance the treatment [1, 5, 20, 21, 32]. Consi-
dering upper limb therapies, a robot can provide assistance
with varying degrees of compensatory movements for the
affected limb. There is evidence that the recovery is more
effective when the patient actively participates in the
training [13]. It is hypothesized that for a successful
rehabilitation, it is crucial to motivate the patient to
actively participate during the training exercises.
T. Nef (&) � M. Mihelj � R. Riener
Sensory-Motor Systems (SMS) Laboratory,
ETH Zurich, TAN E, Tannenstrasse 1,
8092 Zurich, Switzerland
e-mail: [email protected]
URL: www.armin.ethz.ch
T. Nef � M. Mihelj � R. Riener
Spinal Cord Injury Center, University Hospital Balgrist,
University Zurich, Zurich, Switzerland
123
Med Bio Eng Comput
DOI 10.1007/s11517-007-0226-6
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
New actor, sensor, and control strategies can make
robots ‘‘intelligent’’ providing measurement data to assess
the rehabilitation progress, and to gain insight into the
underlying pathology and allowing the patient to actively
participate in the training.
1.2 Rationale for robot-aided arm therapy
One-to-one manually assisted training has several limita-
tions. The training is labour-intensive, time consuming,
and, therefore, expensive. The disadvantageous conse-
quence is that the training sessions are often shorter than
required for an optimal therapeutic outcome. Finally,
manually assisted movement training lacks repeatability
and objective measures of patient performance and
progress.
In contrast, with automated, i.e. robot-assisted, arm
therapy, the duration and number of training sessions can
be increased, while reducing the number of therapists
required per patient. A long-term automated therapy
appears to be the only way to make intensive arm training
affordable for clinical use. As the actual version of the
ARMin-robot is still in the prototype phase, safe operation
requires that one therapist supervises the training (c.f.
2.10). In the future, it will be possible that the patient can
be treated by the device with less supervision. Therefore,
the therapist will be able to manage several robotic devices
or he will be able to do other work besides. Thus, personnel
cost can be reduced. Furthermore, the robot provides
quantitative measures, thus, supporting the observation and
evaluation of the rehabilitation progress.
Several groups have proposed robots to assist physio-
therapy and rehabilitation of the upper limbs (see [26, 27]
for review). The devices provide a varying degree of
assistance to the patient’s movements, ranging from no
assistance if the patient has sufficient voluntary control, to
full assistance, where the patient can behave passively.
New control strategies have been introduced that allow the
machine to comply with forces exerted by the patient
enabling new possibilities for rehabilitation while guaran-
teeing safety for the patient [4, 27, 28].
1.3 Requirements for a rehabilitation robot
In the design and application of rehabilitation robots,
medical aspects must be taken into account to ensure a
successful training. It is crucial that the robot is adaptable
to the human limb in terms of segment lengths, range of
motion and the number of degrees-of-freedom (DOF). A
high number of DOF allows a wide variety of movements,
with many anatomical joint axes involved. However, this
can make the device complex, inconvenient and expensive.
It remains an open issue to assess how many DOF are
optimal for upper limb rehabilitation. The question is
whether therapeutic outcome can be maximized, if the
robot acts on the entire extremity rather than on single
joints only. To answer this question would require a clin-
ical study of an enormous sample size performed with
various devices. However, there is evidence that a therapy
focussing on activities of daily living (ADL) not only
increases the patient’s motivation but also yields an
improved therapeutic outcome, compared to therapies
focussing on single joint movements [2, 22, 25, 34]. To
allow ADL training, a robot must be able to move the
patient’s arm in all relevant degrees of freedom and to
position the human hand at any given point in space. This
can be achieved by an end-effector-based robot or by an
exoskeleton-type device.
End-effector-based robots are connected with the
patient’s hand or forearm at one point. From a mechanical
point of view, these robots are easier to realize and thus,
many research groups work with end-effector-based devi-
ces [7, 9, 16]. In contrast, the structure of exoskeleton
robots resembles the human arm anatomy [29]. Conse-
quently, the arm is attached to the exoskeleton at several
points. Adaptability to different body sizes is easier in an
end-effector-based system, i.e. where the robot moves the
arm by inducing forces only on the patient’s hand. In
contrast, exoskeletal systems are more difficult to adjust,
because each robot link must be adjusted to the corre-
sponding patient arm segment. However, the advantage of
an exoskeleton system as compared to the end-effector-
based approach is that the arm posture is statically fully
determined. Torques applied to each joint can be controlled
separately and hyperextensions can be avoided by
mechanical stops. The possibility to control torques in each
joint separately is essential, e.g. when the subject’s elbow
flexors are spastic. This involuntary muscle activation
results in an increased resistance against movements. To
overcome the resistance, elbow torque up to 20 Nm is
necessary (Table 1). This must not induce any reaction
torques or forces in the shoulder joint, which can be
guaranteed by an exoskeleton robot but not by an end-
effector-based one. This is important because the shoulder
girdle is a rather instable joint and the head of the humerus
bone is held in its position by muscles and tendons and not
by ligaments and bones. If one applies high shear forces to
the shoulder joint, humerus head dislocation can occur.
That’s the reason why therapists use both hands when
they mobilize a spastic elbow joint. With the goal to avoid
exercise forces to the shoulder, one hand holds the lower
arm while the other hand holds the upper arm—comparable
to exoskeleton robots with a cuff fixed to the lower arm and
a cuff fixed to the upper arm.
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
1.4 Patient cooperative arm therapy and specific
research aims
Since the therapy progress depends on training intensity
and training duration, the motivation of the patient turns
out to be a key factor for an efficient rehabilitation [14].
The patient needs to get motivated to contribute actively to
the movement, which may enhance recovery. In addition,
to make longer training duration and more training sessions
possible, it is crucial that the patient enjoys the therapy,
stays concentrated, and does not get bored.
The so called ‘‘cooperative arm therapy’’ discussed in
this paper takes into account the following key aspects that
motivate the patient: (1) the device stimulates the three
most important sensory modalities of the patient, i.e. the
haptic, visual and auditory senses and (2) the robot and
the patient cooperate and interact, i.e. the robot assists the
patient just as much as needed to perform a particular
movement task.
While many clinical studies (see [26] for review) have
been conducted with end-effector-based robots with limi-
ted possibilities to control position and orientation of the
human arm in the three-dimensional space, not much
clinical evidence has been reported from work with actu-
ated arm-exoskeleton robots. The key aspects of this pro-
ject are that the ARMin device allows precise joint
actuation and 3D movement of the arm, that the device
allows to work in ‘‘patient cooperative’’ control modes and
that the device includes a comprehensive audiovisual user
interface.
2 Methods
2.1 Specifications
Training of ADL includes tasks like eating, drinking,
combing hair, etc. For most of these ADL tasks, the hand
has to reach a point in space, grasp an object, and then
control position and orientation of the object until the task
is completed. Therefore, the robot must be able to support
movements of the shoulder, the elbow, and the wrist.
Approximating the shoulder by a three-DOF ball-and-
socket joint, and allowing elbow flexion/extension, pro/
supination of the lower arm and wrist flexion/extension,
results in a device with at least six active DOF. To simplify
the task, our first prototype was built only with four active
DOF supporting the movements of the shoulder joint and
elbow flexion/extension.
The range of motion (ROM) must match as close as
possible the ROM of the human arm [35]. In order to obtain
a satisfactory control performance of patient-cooperative
control strategies, which are based on impedance and
admittance architectures, the robot must have low inertia,
low friction and negligible backlash. Furthermore, the
motor/gear unit needs to be back drivable. Back-drive-
ability is required for good performance of the impedance
control [10–12] and it is advantageous for the safety of
exoskeleton robots (cf. 2.10).
The required velocities and accelerations have been
determined by measuring the movements of a healthy
subject during two ADL tasks (eating soup and manipu-
lating a coffee cup). These values served as inputs for a
simple dynamic model applied to estimate the required
joint torques. In order to ensure that the robot will be strong
enough to overcome resistance from the human against
movements due to spasms and other complications that are
difficult to model, rather high values have been selected
(Table 1). The required end-point payload is 1 kg and end-
point position repeatability is 10 mm. These values allow
manipulation of objects like a coffee cup.
Furthermore, it is required that the robot is easy to
handle and that safety is always guaranteed for both patient
and therapist.
2.2 Kinematics
A semi-exoskeleton solution has been selected for the
mechanical structure of the robot called ARMin (Fig. 1).
The robot is fixed via an aluminium frame at the wall with
the patient sitting in a wheelchair, placed beneath. The
patient’s torso is fixed to the wheelchair with straps and
bands (Fig. 8). ARMin comprises four active and two
passive DOF in order to enable elbow flexion/extension
and spatial shoulder movements [24]. The distal part is
characterized by an exoskeleton structure, with the
patient’s lower and upper arm placed inside orthotic shells.
Table 1 Requirements for the range of motion (ROM), velocity and the maximal torques
Axis ROM Velocity (�/s) Acceleration (�/s2) Torque (Nm)
Axis 1 (Vertical shoulder rotation) 100� to �135� 71 103 20
Axis 2 (Horizontal shoulder rotation) 135� to �45� 60 129 20
Axis 3 (Internal/external shoulder rotation) �50� to 95� 150 245 10
Axis 4 (Elbow flexion/extension) 0� to 135� 91 116 20
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
The exoskeletal part drives the internal/external rotation of
the upper arm and the elbow joint, whereas horizontal and
vertical shoulder rotation is actuated by an end-effector-
based part connecting the upper arm with the wall-mounted
axes 1 and 2.
The robot becomes statically determined only in com-
bination with the human arm. To prevent the robot from
falling down and to avoid unfavourable shear forces and
torques onto the human shoulder, the weight of the robot is
compensated by a passive counterweight. This is achieved
by a counterweight of 2.5 kg that is connected to the robot
via a cable guided by two pulleys.
This end-effector-type kinematics for the shoulder joint
has been selected as exoskeleton mechanics would make it
difficult to get the robot axes in alignment with the ana-
tomical axes of the human, and misalignments would
mechanically overstress and potentially harm the shoulder.
2.3 Mechanics and actuation
The vertically oriented linear motion module (axis 1)
drives vertical shoulder rotation (flexion/extension and
abduction/adduction) movements. Horizontal shoulder
rotation (horizontal flexion/extension and horizontal
abduction/adduction) is realized by a backlash-free and
back-drivable harmonic drive module attached to the slide
of the linear motion module (Table 2).
The interconnection module connects the horizontal
shoulder rotation drive with the upper arm rotary module
via two hinge bearings. Elbow flexion/extension is realized
by a harmonic drive rotary module.
Internal/external shoulder rotation is achieved by a
special custom-made upper arm rotary module that is
connected to the upper arm via an orthotic shell. For easy
access to the patient’s arm, the module is made out of two
half-cylinders. An inner half-cylinder (Fig. 2) is guided by
32 ball bearings fixed to the exterior wall. It is actuated by
three steel cables fixed to the two ends of the cylinder and
rolled around the extension of the motor shaft. This guid-
ance allows transfer of static loads in several DOF while
remaining backlash-free and enabling low friction circular
motion. Custom-made cuffs (Fig. 3) ensure a comfortable
fixation of the patient to the arm rotary module.
2.4 Adaptation to different body sizes
To apply the robot to patients of different sizes, the lengths
of exoskeleton segments need to be adjustable. Table 3 and
Fig. 4 show the ranges of all possible adaptations.
In order to fix the shoulder position of the patient under
the fulcrum at axis 2 (horizontal shoulder rotation), an
individually adjustable belt system was constructed
(Fig. 9). An optional hand support (Fig. 4) can be added to
allow comfortable hand posture.
2.5 Sensors and control hardware
The four exoskeleton actuators are equipped with optical
incremental sensors and redundant potentiometer-based
sensors for position measurements. A six-DOF force-tor-
que sensor beneath the horizontal arm rotation module
measures forces and torques of the shoulder joint (axis 1–
3). The elbow torque is measured by a separate torque
sensor (Fig. 1).
The controller runs on a Matlab/Simulink XPC target
(The Mathworks, Inc., USA) computer with 1 ms loop
time. Four analogue channels provide inputs for the current
amplifiers (Maxon 4-Q-DC Servoamplifier ADS 50/5;
Maxon Motor AG, Switzerland). The four-encoder signals,
the analogue signals from the redundant potentiometer-
Fig. 1 Mechanical structure of ARMin [24]
Table 2 Actuation of the four
axesAxis Gear Motor type
Axis 1: Vertical shoulder rotation Ball screw, 10 mm/rot Maxon RE40, brushed DC
Axis 2: Horizontal shoulder rotation Harmonic drive 1:100 Maxon RE35, brushed DC
Axis 3: Internal/external shoulder rotation Cable drive 1:24.5 Maxon RE40, brushed DC
Axis 4: Elbow flexion/extension Harmonic drive 1:100 Maxon RE35, brushed DC
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
based position sensors are interfaced to the Sensoray 626 I/
O (Sensoray Company Inc., USA) interface card.
The graphical user interface runs on a computer with
Windows operating system (Microsoft Corporation, USA)
and is connected with the real time target by a local area
network using TCP/IP protocol.
Two graphical displays are used: (1) the therapist robot
interface (TRI)—a standard LCD monitor display, and (2)
the patient robot interface (PRI)—a double-projector sys-
tem with polarized images to generate graphical 3D
scenarios for the patient. The TRI contains a dialog-based
graphical user interface, which allows the therapist to enter
the patient’s data, choose the therapy mode, display the
actual patient performance and save the therapy log file.
These data are not shown to the patient, who, instead, looks
onto the screen of the PRI, where the tasks are displayed by
animated 3D scenarios (Fig. 6).
2.6 Dynamic model and basic control issues
A dynamic model of the robot and the human arm has been
developed to simulate and evaluate different control
schemes. An inverse dynamic and a direct dynamic model
have been derived using Lagrange methodology. The direct
dynamic model is given by
€q ¼ M�1ðqÞðs� Cðq; _qÞ _q� GðqÞÞ ð1Þ
where M is the inertia matrix, Cðq; _qÞ _q is the vector of
Coriolis and centrifugal torques, GðqÞ is the vector of
gravity torques and q is the vector of joint positions. The
elements of the robot have been modelled in 3D and the
matrices of inertia and the centres of gravity have been
calculated using the CAD numerical finite element calcu-
lation (AutoCAD, Autodesk, USA). The upper arm and the
lower segments of the human arm have been modelled as
conical frustums with homogeneous mass, equal to that of
water [8].
The inverse dynamic model can be expressed by
s ¼ MðqÞ€qþ Cðq; _qÞ _qþ GðqÞ ð2Þ
where s is the vector of joint torques. In general, the
following torque contributions appear [31]
Fig. 2 Upper arm rotary module for internal/external shoulder
rotation (opened to allow view inside)
Fig. 3 Cuff that connects the rotary module with the human upper
arm
Table 3 Ranges of the adaptations to different body sizes
Label Description Range (cm)
a Shoulder height of the sitting patient 90–110
b Upper arm length 27–42
c/d Lower arm length 20–32
e Wrist circumference 16–24
F Upper arm circumference 20–40
G Optional hand support –
Fig. 4 Six possibilities to adapt the robot to different body sizes
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
si ¼ smi � sfi � shi ði ¼ 1; . . .; 4Þ ð3Þ
where si is the driving torque of joint i, smi is the torque of
joint i delivered by the motor, sfi is the torque of joint i due
to joint friction and shi is the torque of joint i caused by the
human. The velocity dependent part of friction has been
identified and compensated.
A PD controller with a computed torque feed forward
portion has been implemented for position control and
evaluated by simulation (Fig. 5). Three different strategies
have been compared regarding their control error and the
computational effort with regard to the number of opera-
tions needed per time step. Sample time was fixed to 1 ms.
In the first strategy, the PD values where selected by
simulation and remained constant while the feed forward
part was calculated by the inverse dynamic approach (Eq.
2) (computed torque control). In the second strategy, the
feed forward part was simplified to the gravity part leading
to
sff ¼ GðqÞ ð4Þ
And in the third strategy the feed forward part has been
set equal to sff ¼ 0, which reduces the strategy to a simple
PD controller. Sinusoidal trajectories have been used for
these experiments, while in the real application the tra-
jectory generation described in Chap. 2.7 has been used.
The model served also to investigate stability of the robot
interacting with the human. For this purpose, a human that
produces perturbations of different frequencies and
amplitudes has been simulated.
2.7 Trajectory generation for the mobilisation therapy
Mobilisation therapy is based on repetitive movements of
the human arm on a patient-specific trajectory while the
patient remains passive. This can prevent joint degenera-
tion, preserve the patient’s mobility and joint flexibility,
and reduce spasticity. The mobilisation therapy is executed
in two steps. First, the therapist moves the patient’s arm
together with the robot on a desired trajectory. The exact
shape of the movement (ranges, speed) can be solely
defined by the therapist taking into consideration the
individual impairment of the patient. During the recording
phase, the robot’s gravity and friction are compensated so
that the therapist only feels the forces and torques neces-
sary for moving the human arm. Once the trajectory is
recorded, the relevant points of the trajectory are deter-
mined in order to enable calculation of a smooth and
‘‘human-like’’ movement path. This is required as the tra-
jectory performed by the therapist is often shaky. Simple
low-pass filtering is not appropriate, as it would cut the
movement at the extremes. In the second step, the robot
repeats the trajectory with an adjustable velocity.
Jung-Hoon et al. [15] developed a simple method to
extract a small number of way-points for the trajectory of a
mobile robot. This method has been adapted in the way that
turning points of the recorded position data are detected
and used as via points. Once these points are detected, a
smooth trajectory connecting them is generated by a mini-
mum jerk approach [6]. This approach is quite common to
describe the kinematics of smooth human arm reaching
movements. The resulting trajectory serves as input for the
position controller (c.f. 2.6).
2.8 Audiovisual display
An appropriate audiovisual display is imperative for the
game supported therapy (cf. 2.9) where movement tasks
need to be displayed to the patient. The mobilisation
therapy (cf. 2.6) could also be performed without any
audiovisual feedback. However, showing a virtual arm can
help the patient to realize his or her arm posture.
As the robot is able to perform ADL-related movements
in space, a stereographic display is used to present virtual
objects in three dimensions. The screen is rather large
(2 m · 2.7 m) so that even patients with mild visual
impairments can recognize the scenarios and tasks, and
Fig. 5 Computed torque
position control loop for the
mobilisation therapy
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
generate a feeling of presence. Additional sound can help
to increase the feeling of presence during the game sup-
ported therapy.
Two projectors with polarizing filters are used to project
stereographic images via back projection onto the screen.
Patients wear lightweight passive polarizing glasses to
see 3D images. The Open Inventor clone Coin 3D
(http://www.coin3D.org) has been used as a scene graph. In
combination with the quad buffer of the NVIDIA Quadro
Fx 4000 graphic card (NVIDIA Corporation, USA) ste-
reoscopic images are generated without any additional
programming effort.
The virtual environment (VE) consists of a graphical, an
acoustic, and a physical model of the scene. The physical
model is required to determine force feedback signals in
order to allow the patient to physically interact with the
VE. It is implemented on the real time computer (XPC
Target) and updated with a sampling rate of 1 kHz. The
graphical model is implemented with Coin 3D and runs on
a Windows computer with a sampling rate of approxi-
mately 100 Hz.
Several different scenarios have been implemented.
During the mobilisation therapy, an avatar figure showing
the actual posture of the patient’s arm is presented (Fig. 6
upper left). A ball game scenario (Fig. 6 upper right)
includes a ball rolling down an inclined table and a hand
connected to the handle. The ball is reflected by the walls
and the patient’s task is to catch the ball with the handle.
The colour of the handle changes depending on the per-
formance of the patient (Fig. 11). The sounds of the rolling
ball and the collisions with the wall and the handle are
displayed to increase the level of realism. A labyrinth game
scenario has been implemented (Fig. 6, lower left) with a
red ball indicating the cursor that moves according to the
patient’s spatial hand movement. Force feedback is pro-
vided whenever the cursor touches the wall of the laby-
rinth. To motivate the patient, objects can be collected on
the way from the bottom to the top of the labyrinth.
In the ADL training mode, the patient is asked to grasp
an object by approaching it with the virtual hand and to put
it onto the table (Fig. 6, lower right). The control strategy
for these kinds of tasks is based on the minimal interven-
tion principle, which allows an efficient exploitation of task
space redundancies and results in user-driven movement
trajectories [23]. The basic idea of this approach is that
deviations from the trajectory are corrected only when they
interfere with the task performance. In other words: if the
patient is doing ‘‘better’’ than the robot, then the robot lets
the patient do so.
2.9 Control scheme for the ball game
A simple ball game has been implemented to demonstrate
and validate the concept of game-supported therapy. The
task is to catch a virtual ball rolling down an inclined
virtual table (Fig. 6, upper right). The ball movement in y
direction (vertical direction on the screen) is governed by
Newton’s law:
vyball¼ v0 � g sinðaÞ ð5Þ
where v0 is the initial velocity of the ball, g is the gravity
acceleration and a is the inclination of the virtual table. The
initial velocity v0 and inclination a can be adjusted for each
Fig. 6 Scenarios for the arm
therapy. Upper left Movement
therapy. Upper right Game
supported therapy (ball game).
Lower left Game supported
therapy (labyrinth game). Lowerright ADL training scenario
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
patient in order to increase or decrease the difficulty of the
game.
The control strategy (Fig. 7) is based on an impedance
controller, where the assisting force Fxref is determined,
based on the relative distance between the ball and the
human hand position. The force in horizontal direction
Fxref, that the robot exerts onto the patient’s arm to push
him towards the ball position is:
Fxref ¼0
K xball� xhandð Þ � ð1� yballÞ�Bdxhand
dt
�if yball [0:5if yball�0:5
ð6Þ
K is a constant value that can be adjusted by the therapist
and B is a damping factor. Typical values are K = 10 N/m
and B = 0.01 Ns/m.
During the first part of the rolling-down sequence
(yball > 0.5), the force Fxref is always zero. This kind of
delay gives the patient time to try to bring his hand to
the ball position without robotic support. In the second
part of the sequence (yball � 0.5), the force Fxref is
proportional to the distance between the hand and the
ball (xball–xhand) and to the factor (1–yball), and therefore
zero if the patient was able to bring his hand to the ball
position during the first sequence. If not, then the sup-
porting force rises when the ball approaches the zero
level (yball ? 0).
2.10 Passive and active safety
Safety was a main issue during the design of the robot.
Passive safety features (no sharp edges, mechanical end
stops to guarantee that no joint can exceed the anatomical
range of motion, etc.) are combined with active safety
features. Four redundant absolute position-sensing poten-
tiometers—one for each joint—allow detecting malfunc-
tion of a position sensor. Several surveillance routines are
implemented in the software. These include current and
speed monitoring, a robot self-collision detection algorithm
and several watchdog systems.
Whenever an abnormal event is detected, the safety
circuit immediately cuts the power of the motor drives. As
the robot is designed with a passive weight compensation
system (pulley, rope, and counterweight, see Fig. 8) it does
not collapse after power loss. Since all drives are back
drivable, the robot can easily be moved manually by a
therapist in order to release the patient from a potentially
uncomfortable or dangerous posture.
Last but not least, the physiotherapist always observes
the training holding a deadman button in his hand.
Releasing the button interrupts the motor power and stops
the robot immediately. This can also be achieved by
pressing one of the three emergency stop buttons. It is
expected that future robots will not require a permanent
supervision and that the deadman button could be omitted.
Beside patient safety, also safety of the therapist needs
to be considered. As the robot does not know the position
of the therapist, it is important that the therapist is aware of
the danger of collisions with the robot. Nevertheless, the
Human
Robot
q
q
τ
hτ
Gravity compensation
gravτ
Friction compensation
frictτ
F/T Sensor
TJP
Bs
+
+ ++
++ +
refxF
xf−
Kbally
Physical model of the inclined
plane and of the ball
+ −
bally
ballx
handx
)0( handx
)(ballball yx
visual information
graphical display
x
y
K+
0 5.0 1
1
constant gain
TJ −)(qk
direct kinematics
−
Fig. 7 Control strategy and
signal flow of the ball game
Fig. 8 ARMin prototype that has been used for all experiments
described in this paper
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
probability of a severe accident is low because of the fact
that the maximal speed of the robot is limited by the sur-
veillance circuit. Furthermore, the therapist can always
interrupt the motion by releasing the deadman button. A
detailed risk analysis shows that the risk for a patient and a
therapist using the robot is acceptable with respect to the
expected rehabilitative benefit to the patient.
3 Results
ARMin has been installed at Balgrist University Hospital
in Zurich, Switzerland (Fig. 8). After approval of the ethics
committee of Zurich, the device first has been tested with
healthy subjects. Then, a pilot study including eight
hemiplegic and three incomplete spinal cord injured sub-
jects has been carried out, followed by three clinical single-
case studies with chronic stroke patients performing
intensive robotic training of longer duration. The tests with
healthy subjects and the pilot study served to prove the
functionality of the device, without looking at possible
improvements in the motor performance of the patients.
Possible improvements have then been assessed by the
clinical studies with the three chronic stroke patients
(publication in preparation).
3.1 Control scheme and stability
The dynamic model was used to simulate the overall system
and to evaluate the performance of three different control-
lers: Computed torque, PD in combination with gravity
compensation, and PD alone. Sinusoidal reference trajec-
tories (T = 2ps) where selected as reference trajectories for
all four axes. The mean position error was calculated for
every axis and for all axes together. Three different condi-
tions where selected. First, the ideal model was used, where
ideal means that the plant model (direct dynamic model of
robot and human) was equal to the model used for feed
forward torque calculation (inverse dynamic model).
Second, the ideal model was used with sensor discretisation.
As the encoders deliver a digital signal, no noise was
introduced. And third, the direct dynamic model was fal-
sified by increasing the mass of the human arm by 30%,
simulating modelling errors. The PD-values where for all
cases the same and they have been determined by simula-
tions using the dynamic model (Axis 1: P = 10,000 N/cm,
D = 50 Ns/cm; Axis 2: P = 1,000 Nm/rad, D = 10 Nms/
rad; Axis 3: P = 1,000 Nm/rad, D = 10 Nms/rad; Axis 4:
P = 300 Nm/rad, D = 5 Nms/rad). For all simulations,
passive behaviour of the human was assumed. Table 4
summarises the results.
The third condition reflects reality best. In this condi-
tion, the computed torque controller does not show sig-
nificantly better results than the PD controller with gravity
compensation, although the computational effort was much
higher. In contrast, the pure PD controller required only a
few operations, whereas it showed large errors. As a
compromise, the PD controller with gravity compensation
was chosen and implemented in the robot controller
hardware.
In order to find out whether the controller stays stable
when the user applies rhythmic disturbances, sinusoidal
disturbances acting onto one axis of the robot have been
simulated and the position and the velocity error of the
simulated robot have been observed. The frequencies of the
Table 4 Mean position errors of different control schemas
PD +
computed
torque
PD + gravity
compensation
PD
Number of operations
per step
8000 150 20
Position error with ideal
plant model
0.00007� 0.06� 0.5�
Position error with ideal plant
model with sensor discretisation
0.007� 0.06� 0.5�
Position error with falsified plant
model with sensor discretisation
0.05� 0.08� 0.5�
05
1015
2025
30 0
5
10
15
20
25
30
0
0.1
0.2
0.3
0.4
Amplitud
e [Nm]
Frequency [Hz]
Pos
ition
err
or [
rad]
0
10
20
30 0
10
20
300
0.5
1
1.5
2
2.5
3
Amplitude [Nm]
Frequency [Hz]
Vel
ocity
err
or [
rad/
s]
Fig. 9 Left Simulated position
error with sinusoidal
disturbances produced by the
human. Right Simulated
velocity error with sinusoidal
disturbances produced by the
human. The figure shows when
the simulated robot becomes
marginally stable, which means
that the output is still bounded
but vibrations occur
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
disturbances varied from 0 to 30 Hz with amplitudes
varying from 0 to 30 Nm, and the disturbances have been
injected at the would point where the user exercises forces
onto the robot. For this simulation, the saturation of the
motors has been set to 25 Nm.
A series of simulation studies leads to the result that the
position error stays small while the velocity error varies with
the frequency and amplitude of the disturbance (Fig. 9).
3.2 Functional tests with healthy subjects
Within the tests with healthy subjects, the ROM, the con-
trol and the overall functionality have been validated
(Table 5). These tests included 20 min’ mobilisation ther-
apy and 40 min’ game therapy.
3.3 Pilot study with patients
A pilot study was carried out to validate the patient comfort
and handling, i.e. whether the robot is adjustable to dif-
ferent patient sizes, whether the cuffs are comfortable for
the patients, and whether the patients are able to perform
the movement tasks. Furthermore, the patient acceptance
and motivation were interrogated.
Using the same protocol as for healthy subjects, the
robot has been tested with 11 patients for a cumulative
duration of more than 76 h. The robot could easily
accommodate subjects with body sizes between 155 and
192 cm. The fixation of a patient in the robot takes
approximately 5 min.
Figure 10 shows an example for a trajectory recorded by
a therapist and the resulting trajectory after applying the
minimum jerk approach. Note that the extreme values of
both trajectories correspond with each other.
Figure 11 shows the results of the game therapy with
two different hemiplegic patients. A ball has been dis-
played falling down on the graphical display. The patient
had to catch it by moving the virtual hand (Fig. 6, upper
right). Subject A was able to catch the ball mostly without
any robot support. The robot assisted patient B by pushing
the patient with the force Fxref towards the ball (Fig. 11), if
the patient was not able to catch the ball himself (cf. Chapt.
2.9). Note that the robot support is always delayed rela-
tively to the change in the horizontal ball position. In this
Table 5 Technical data of the device
Payload (Endpoint) 2 kg
Weight (excl. controller hardware, frame), approx. 6.5 kg
Repeatability (endpoint) ±3 mm
Stiffness (endpoint)a >714 N/m
Minimal apparent mass (endpoint) 240 g
Bandwidth for small endpoint movements (0.1 m),
according to Ellis et al. [3]
2.1 Hz
Axis data (subject body size 181cm): Range: Precision: Stiffness of the
motor/gear unit:
Max
torquesbMobilisation
torquesc
Axis 1 (Vertical shoulder rotation) 44� to �59� 0.1� 136 kN/m 57 Nm 10.15 Nm
Axis 2 (Horizontal shoulder rotation) 130� to �50� 0.1� >1,600 Nm/rad 52 Nm 9.36 Nm
Axis 3 (Internal/external shoulder rotation) �60� to 95� 0.5� 229 Nm/rad 11 Nm Not measured
Axis 4 (Elbow flexion/extension) 5� to 119� 0.1� >1600 Nm/rad 27 Nm 24.36 Nm
a Stiffness measured at the endpoint by applying 20 N while the motors are position controlled and a dummy arm was fixed in the deviceb Torques measured during maximum voluntary contractions of healthy personsc Torque values necessary for mobilisation. Maximum values that occurred during 60 mobilisation movements carried out on three stroke
patients
Fig. 10 Recorded trajectories during teach mode and the derived
minimized jerk trajectory
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
way, the subject is first encouraged to voluntarily move the
virtual hand toward the target. Only when the height of the
ball decreases to a level that would not allow the subject to
catch the ball, the robot support is activated and guides the
arm to the target position.
An important goal was to assess the patient motivation
for this kind of therapy. Thus, five patients had to fill out a
questionnaire after the therapy sessions. As most of them
where not able to read and write, they where interrogated
by an independent person. Five questions where asked and
the patients could answer with a number between 1 and 6,
where 1 means ‘‘not at all’’ and 6 means ‘‘very much’’. This
grading system has been selected because it is well known
among the patients as it is adapted from the local school
system (Table 6).
The pilot study included mobilisation therapy and ball
game therapy. The labyrinth game and the ADL training
scenarios have been tested with some healthy subjects, but
have not been evaluated with patients yet.
4 Discussion
4.1 Control scheme and stability
It has been shown that the computed torque controller
allows a precise position control in the case of an ideal
plant. The remaining error is due to the numerical impre-
cision of the calculations. In the realistic scenario, with
discrete sensor information and modelling error, the dif-
ferences between computed torque and gravity compensa-
tion with PD controller become smaller and therefore, the
gravity compensation with PD controller is a good com-
promise between calculation time and precision. The
measured position error of 0.1� for axis 1, 2 and 4 and 0.5�for axis 3 is close to the values determined by the simu-
lations and satisfactory for rehabilitation purposes.
Figure 9 shows the position error and the velocity error
when the human acts with sinusoidal disturbances at dif-
ferent frequencies. The position error does remain rather
small, except when the amplitude gains values above
25 Nm. The reason for this clear increase is that in the
simulations the actuator torques were limited to 25 Nm.
The velocity error shows more variations. These variations
became indeed observable in small, high frequency vibra-
tions that were not visible but produced an audible noise.
180 185 190 195 200 205 210 215 220 225 230-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.81
Time [s]N
orm
alis
ed le
vels
Subject A
xball
xhand
Fxref
320 325 330 335 340 345 350 355 360 365 370-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Time [s]
Nor
mal
ised
leve
ls
Subject B
xball
xhand
Fxref
Fig. 11 Movement and force
support recorded from two
different hemiplecic subjects
playing the ball game. Subject
A (upper graph) does not need
support from the robot, while
subject B (lower graph) needs
some support
Table 6 Questionnaire
Question (translated from German) Average
mark
(n = 5)
1. Was the fixation in the robot comfortable? 4.6
2. Did you like to do the mobilisation therapy? 5.25
4. Did you like to play the ball game? 5.0
3. Do you think that this sort of therapy could
help you to regain motor function?
5.25
5. What was your overall impression from the device? 5.5
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
Many hours of operation with adjusted PD controller
parameters without a patient did not yield instable behav-
iour. Only high-frequency and high-amplitude external
perturbations (e.g. induced by the patient) cannot be
compensated by the controller and can cause vibrations
(Fig. 9, right).
However, assuming that the bandwidth of human arm
movement is limited to 10 Hz [30], such high-frequency
and high-amplitude perturbation cannot be induced by the
human.
This is not a strict stability proof, but both, the simu-
lation and the experimental results, support the hypothesis
that the robot stays stable when it interacts with the human
arm.
4.2 Specifications of the device
Most of the specifications for the robot (Table 1) could be
fulfilled (Table 5). The ROM of the vertical shoulder
rotation is limited by the mechanical construction of the
robot.
4.3 Pilot study
The teach-and-repeat mode allows the therapist to easily
define an appropriate movement and deliver smooth
movements. While the patient is passive during the mo-
bilisation therapy, he or she is active during the ball-game-
supported therapy. The results of the questionnaire suggest
that the patients like both, the ball game and the mobili-
sation. The patients rated their fixation in the device with
the average mark 4.6, which is significantly lower than the
other marks. As the elbow fixation did not cause any
problems, this relatively bad mark seems to be related to an
uncomfortable shoulder fixation.
In fact, the shoulder actuation module is the most critical
component (Fig. 1). It includes 3 actuated DOF and 2
passive DOF, which is advantageous in that the human
shoulder is not constrained by the robot fixation. This
allows movements of the shoulder in 3 rotary DOF with
large ROM. Although the weight of the robot has been
compensated by a passive counterbalance system, addi-
tional subject-dependent weight and active joint torques
produced by the robot will be transferred to the human
shoulder as the human arm is closing the open kinematic
chain of the robot. During standstill, the distal part of the
robot and the human arm beyond the passive axis 5 (Fig. 1)
is in a static equilibrium about axis 5. Therefore, the
resulting vertical force acting on the shoulder depends on
the robot and arm position and the weight of the mechani-
cal components and the human arm. The weight of the
distal part of the robot alone induces a vertical shoulder
force of approximately 8.5 N (robot arm in a horizontal
position, elbow extended). Furthermore, the drives of axes
1–3 produce reaction forces that are transmitted through the
shoulder joint to the environment, which further increases
the load at the shoulder.
Therefore, patients with instable shoulder joints, e.g.
resulting from shoulder subluxations, could not be treated
by the robot. Thus, our shoulder actuation approach is
acceptable for healthy subjects but problematic for patients
with shoulder problems. As many stroke patients suffer
from shoulder problems, it has been suggested to modify
the shoulder actuation module.
4.4 Prototypes
All simulations and measurements presented in this paper
have been performed with the presented 4 DOF version of
ARMin (Fig. 8). Derived from the results of this work, a
new version with 6 DOF has been set up (Fig. 12). The
robot is statically determined and equipped with a new
shoulder actuation principle. A system of linkages moves
the centre of rotation of the shoulder in the vertical plane,
when the arm is lifted. This function is required to provide
an anatomically correct shoulder movement and to avoid
the shoulder getting mechanically overstressed due to
misalignment of the technical and anatomical joint axes
when lifting the upper arm above face level. The device is
not evaluated yet.
Fig. 12 ARMin II—New version with 6DOF and statically deter-
mined shoulder actuation
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
5 Conclusion and outlook
A new semi-exoskeleton arm rehabilitation robot called
ARMin has been developed and tested with numerous
stroke and SCI patients. Providing movements in four
DOF, ARMin can support spatial arm movements. A
patient-cooperative control strategy has been presented and
evaluated. It supports the patient only when necessary. In
combination with an audiovisual display ARMin can be
used to play ball games or to train ADL-related tasks.
These methods have the chance to provoke the patients to
actively contribute to the movement, increase the motiva-
tion of the patient, enhance the intensity of the training and,
thus, improve the therapeutic outcome of the therapy
compared to conventional training methods. Future tech-
nical work will focus on the evaluation of the new shoulder
actuation method and the implementation and evaluation of
further ADL training tasks. Future clinical work will focus
on clinical studies measuring the rehabilitation benefit for
the patient. It needs to be demonstrated that the therapeutic
outcome of arm therapy with the ARMin robot is higher
than the arm therapy delivered by robots with simpler
mechanics.
Acknowledgments This study was supported in part by the NCCR
for Neuroplasticity and Repair, Project 8, Switzerland. We thank
Dr. Gery Colombo from Hocoma AG, Volketswil, Switzerland for
his contribution to this work. We also thank the occupational ther-
apists and Prof. Dr. V. Dietz of the Balgrist University Hospital,
Zurich, as well as Raphael Suard, Stephane Kuhne, Christina Perndl,
Frauke Oldewurtel and Gabriela Kiefer for their contributions to this
work.
References
1. Aisen ML, Krebs HI, Hogan N, McDowell F, Volpe BT (1997)
The effect of robot-assisted therapy and rehabilitive training on
motor recovery following stroke. Arch Neurol 54:443–446
2. Bayona NA, Bitensky J, Salter K, Teasell R (2005) The role of
task-specific training in rehabilitation therapies. Top Stroke
Rehabil 12:58–65
3. Ellis RE, Ismaeil OM, Lipsett M (1996) Design and evaluation of
a high-performance haptic interface. Robotica 14:321–327
4. Emken JL, Bobrow JE, Reinjkensmeyer DJ (2005) Robotic
movement training as an optimization problem: designing a
controller that can assist only as needed. In: Proceedings of
ICORR 2005, 9th international conference rehabilitation robotics,
pp 307–312
5. Finley MA, Fasoli SE, Dipietro L, Ohlhoff J, Macclellan L,
Meister C et al (2005) Short-duration robotic therapy in stroke
patients with severe upper-limb motor impairment. J Rehab Res
Dev 42(5):683–691
6. Flash T, Hogan N (1985) The coordination of arm movements: an
experimentally confirmed mathematical model. J Neurosci
5(7):1688–1703
7. Harwin W, Loureiro R, Amirabdollahian F, Taylor M, Johnson G,
Stokes E, Coote S, Topping M, Collin C et al (2001) In: Marincek
C et al (eds) The Gentle/s project: a new method for delivering
neuro-rehabilitation, asistive technology-added value to the
quality of life AAATE’01.ISO Press, Amsterdam, pp 36–41
8. Herzog W, Nigg BM (1994) Biomechanics of the Muskulo-
skeletal system. Wiley, Chichester
9. Hesse S, Werner C, Pohl M, Rueckriem S, Mehrholz J, Lingnau
ML (2005) Computerized arm training improves the motor con-
trol of the severely affected arm after stroke: a single-blinded
randomized trial in two centers. Stroke 36(9):1960–1966
10. Hogan N (1985a) Impedance control: an approach to manipula-
tion: Part I—theory. ASME J Dyn Syst Meas Control 107
(11):1–7
11. Hogan N (1985b) Impedance control: an approach to manipula-
tion: Part II—implementation. ASME J Dyn Syst Meas Control
107(11):8–16
12. Hogan N (1985c) Impedance control: an approach to manipula-
tion: Part III—applications. ASME J Dyn Syst Meas Control
107(11):17–24
13. Hogan N, Krebs HI, Rohrer B, Palazzolo JJ, Dipietro L, Fasoli
SE, Stein J, Hughes R, Frontera WR, Lynch D, Volpe BT
(2006) Motions or muscles? Some behavioral factors underlying
robotic assistance of motor recovery. J Rehab Res Dev
43(5):605–618
14. Jack D, Boian R, Merians AS, Tremaine M, Burdea GC, Ad-
amovich SV, Recce M, Poizner H (2001) Virtual reality-
enhanced stroke rehabilitation. IEEE Trans Neural Syst Rehab
Eng 9(3):308–318
15. Jung-Hoon H, Ronald CA, Dong-Soo K (2003) Mobile robots at
your fingertip: Bezier curve on-line trajectory generation for
supervisory control. IEEE International conference of intelligent
robots and systems
16. Krebs HI, Hogan N, Aisen ML, Volpe BT (1998) Robot-Aided
Neurorehabilitation. IEEE Trans Rehab Eng 6:75–87
17. Kwakkel G, Wagenaar RC, Koelman TW, Lankhorst GJ, Koet-
sier JC (1997) Effects of intensity of rehabilitation after stroke. A
research synthesis. Stroke 28:1550–1556
18. Kwakkel G, Wagenaar RC, Twisk JWE, Langkhorst GJ, Koetsier
JC (1999) Intersity of leg and arm training after primary middle-
celebral artery stroke: a randomised trial. Lancet 35:191–196
19. Kwakkel G, Kollen BJ, Wagenaar RC (2002) Long therm effects
of upper and lower limb training after stroke: a randomised trial. J
Neurol Neurosurg Psychiat 72:473–479
20. Luft AR, McCombe-Waller S, Whitall J, Forrester LW, Macko R,
Sorkin JD, Schulz JB, Goldberg AP, Hanley DF (2004) Repeti-
tive bilateral arm training and motor cortex activation in chronic
stroke—a randomized controlled trial. JAMA 292:1853–1861
21. Lum PS, Burgar CG, Shor PC, Majmundar M, Van der Loos M
(2002) Robot-assisted movement training compared with con-
ventional therapy techniques for the rehabilitation of upper-limb
motor function after stroke. Arch Phys Med Rehab 83(7):952–
959
22. Langhammer B, Stanghelle JK (2000) Bobath or motor relearning
programme? A comparison of two different approaches of
physiotherapy in stroke rehabilitation: a randomised controlled
study. Clin Rehabil 14:361–369
23. Mihelj M, Nef T, Riener R (2007) A novel paradigm for patient
cooperative control of upper limb rehabilitation robots. Adv
Robot 21(8):843–867
24. Nef T, Colombo G, Riener R (2005) ARMin—robot for move-
ment therapy of the upper extremities. Automatisierungstechnik
53(12):597–606
25. Platz T (2003) Evidenzbasierte Armrehabilitation: Eine system-
atische Literaturubersicht. Nervenarzt 74:841–849
26. Prange GB, Jannink MJ, Groothuis-Oudshoorn CG, Hermens HJ,
IJzerman MJ (2006) Systematic review of the effect of robot-
aided therapy on recovery of the hemiparetic arm after stroke.
J Rehabil Res Dev 43(2):171–184
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900
27. Riener R, Nef T, Colombo G (2005) Robot-aided neurorehabili-
tation for the upper extremities. Med Biol Eng Comput 43:2–10
28. Riener R, Lunenburger L, Jezernik S, Anderschitz M, Colombo
G, Dietz V (2005) Cooperative subject-centered strategies for
robot-aided treadmill training: first experimental results. IEEE
Trans Neural Syst Rehab Eng 13:380–393
29. Sanchez RJ, Liu J, Rao S, Shah P, Smith R, Rahman T, Cramer
SC, Bobrow JE, Reinkensmeyer DJ (2006) Automating arm
movement training following severe stroke: functinal exercise
with quantitative feedback in a gravity-reduced environment.
IEEE Trans Neural Syst Rehabil Eng 14(3):378–389
30. Schouten AC, de Vlugt E, van der Helm FCT, Brouwn GG (2004)
Optimal posture control of a musculo-skeletal arm model. Biol
Cybern 84(2):143–152
31. Siciliano B, Villani L (1999) Robot force control. Kluwer, Boston
32. Stein J, Krebs HI, Frontera WR, Fasoli SE, Hughes R, Hogan N
(2004) Comparison of two techniques of robot-aided upper limb
exercise training after stroke. Am J Phys Med Rehabil 83(9):720–
728
33. Sunderland A, Tinson DJ, Bradley EL, Fletcher D, Langton HR,
Wade DT (1992) Enhanced physical therapy improves recovery
of arm function after stroke. A randomised clinical trial. J Neurol
Neursurg Psychiatry 55:530–535
34. Winstein CJ, Rose DK, Tan SM, Lewthwaite R, Chui HC, Azen
SP (2004) A randomized controlled comparison of upper-
extremity rehabilitation strategies in acute stroke: a pilot study of
immediate and long-term outcomes. Arch Phys Med Rehab
85(4):620–628
35. Winter D (1989) Biomechanics and motor control of human
movement, 2 edn. Wiley, New York
Med Bio Eng Comput
123
Medical and Biological Engineering and Computing, Vol. 45, 2007, p.p. 887-900